[SPARK-8756][SQL] Keep cached information and avoid re-calculating footers in ParquetRelation2 #7154
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -345,39 +345,59 @@ private[sql] class ParquetRelation2( | |
| // Schema of the whole table, including partition columns. | ||
| var schema: StructType = _ | ||
|
|
||
| // Cached leaves | ||
| val cachedLeaves: mutable.Map[Path, FileStatus] = new mutable.HashMap[Path, FileStatus]() | ||
|
|
||
| /** | ||
| * Refreshes `FileStatus`es, footers, partition spec, and table schema. | ||
| */ | ||
| def refresh(): Unit = { | ||
| // Lists `FileStatus`es of all leaf nodes (files) under all base directories. | ||
| // We only care about new FileStatus and updated FileStatus. | ||
| val leaves = cachedLeafStatuses().filter { f => | ||
| (isSummaryFile(f.getPath) || | ||
| !(f.getPath.getName.startsWith("_") || f.getPath.getName.startsWith("."))) && | ||
| (!cachedLeaves.contains(f.getPath) || | ||
| cachedLeaves(f.getPath).getModificationTime < f.getModificationTime) | ||
| }.map { f => | ||
| cachedLeaves += (f.getPath -> f) | ||
| f | ||
| }.toArray | ||
|
Contributor
There was a problem hiding this comment. This is irrelevant to this PR, but I'd like to point out that this check is unnecessary now. IIRC, at the time PR #6012 was merged, |
||
|
|
||
| dataStatuses = leaves.filterNot(f => isSummaryFile(f.getPath)) | ||
| metadataStatuses = leaves.filter(_.getPath.getName == ParquetFileWriter.PARQUET_METADATA_FILE) | ||
| commonMetadataStatuses = | ||
| leaves.filter(_.getPath.getName == ParquetFileWriter.PARQUET_COMMON_METADATA_FILE) | ||
|
|
||
| dataSchema = { | ||
| val dataSchema0 = maybeDataSchema | ||
| .orElse(readSchema()) | ||
| .orElse(maybeMetastoreSchema) | ||
|
|
||
| // If we already have previous loaded schema, merge it with newly loaded schema. | ||
| val dataSchema1 = | ||
| if (dataSchema != null) { | ||
| if (dataSchema0.isDefined) { | ||
| dataSchema0.get.merge(dataSchema) | ||
| } else { | ||
| dataSchema | ||
| } | ||
| } else { | ||
| if (dataSchema0.isDefined) { | ||
| dataSchema0.get | ||
| } else { | ||
| throw new AnalysisException( | ||
| s"Failed to discover schema of Parquet file(s) in the following location(s):\n" + | ||
| paths.mkString("\n\t")) | ||
| } | ||
| } | ||
|
|
||
| // If this Parquet relation is converted from a Hive Metastore table, must reconcile case | ||
| // case insensitivity issue and possible schema mismatch (probably caused by schema | ||
| // evolution). | ||
| maybeMetastoreSchema | ||
| .map(ParquetRelation2.mergeMetastoreParquetSchema(_, dataSchema1)) | ||
| .getOrElse(dataSchema1) | ||
| } | ||
| } | ||
|
|
||
|
|
@@ -426,11 +446,15 @@ private[sql] class ParquetRelation2( | |
| } | ||
|
|
||
| assert( | ||
| filesToTouch.nonEmpty || maybeDataSchema.isDefined || maybeMetastoreSchema.isDefined || | ||
| dataSchema != null, "No predefined schema found, " + | ||
|
||
| s"and no Parquet data files or summary files found under ${paths.mkString(", ")}.") | ||
|
||
|
|
||
| if (filesToTouch.nonEmpty) { | ||
| ParquetRelation2.mergeSchemasInParallel(filesToTouch, sqlContext) | ||
| } else { | ||
| None | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Unfortunately there's still a bug here :) We also need to remove those files that only exist in the old cache (namely removed files). This happens when an existing directory is overwritten.
I think we can first keep both the old cache and the result of
cachedLeafStatuses(), then filter out updated and new files, and at last update the oldFileStatuscache.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I guess you are trying to only keep
FileStatuses of those files that need to be touched during schema merging here. Actually theFileStatuscache must be consistent with the files stored on the file system. Because we also inject the cache toParquetInputFormatinParquetRelation2.buildScanto avoid callinglistStatusrepeatedly there.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hmm, I am thinking, if we only read the footers of the updated and newly added files, the merged schema may be incorrect? For the removed files, it is the same situation. If we don't re-merge all footers' schema, the schema should not be correct.
So this pr should check if we need to re-read all footers and merge schema based on whether the
FileStatuses are updated or not.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is a good point. After reading footers and merging schemas of of new files and updated files, we also need to merge the result schema with the old schema, because some columns may be missing in new files and/or updated files.
Actually I found it might be difficult to define the "correctness" of the merged schema. Take the following scenario as an example:
Initially there is file
f0, which comes with a single columnc0.Merged schema:
c0File
f1is added, which contains a single conlumnc1Merged schema:
c0,c1Removing
f0Which is the "correct" merged schema now?
a.
c0,c1b.
c1I tend to use (a), because removing existing columns can be dangerous, and may confuse down stream systems. But currently Spark SQL uses (b). Also, we need to take metastore schema into account for Parquet relations converted from metastore Parquet tables.
I think this issue is too complicated to be fixed in this PR. I agree with you that we should keep this PR simple and just re-read all the footers for now. It's already strictly better than the current implementation, not mention that schema merging has been significantly accelerated by #7396.